My Bitcoin Trading Journey and Building a Neural Network for Price Prediction

Introduction

My journey into Bitcoin trading began with buying one-tenth of a Bitcoin before the ETFs were announced last year, around November. I’d been watching BTC for some time it but wasn’t until the investment company, BlackRock, started buying and rumours of the ETF started that I finally brought some at $40,000. I watched as the price surged to around $74,000, and I then finally panic sold when Bitcoin hit around $60,000 after it dropped for the fourth time! Despite its initial spike, Bitcoin never quite maintained the momentum. This experience opened my eyes to the volatility and manipulation within the Bitcoin market, leading me to where I am now—developing a neural network to predict Bitcoin’s price movements more accurately.

Learning the Market

What became clear to me was how manipulated the market is, with prices constantly pumping up and down due to the influence of futures trading. Most of the Bitcoin being bought and sold is actually loaned from exchanges, with traders using leverage—often as high as 50 to 1. This means for every £1 a trader puts down, the exchange loans them up to £50 or more. After losing money with the higher leverage, I shifted to a 5 to 1 strategy, and I began to see some success. While I believe this method can outperform traditional stock market returns, I realized that trading based solely on intuition was risky.

The Idea of Building a Bitcoin Prediction AI

Observing the market’s behaviour, I noticed that traders often rely on charts, technical indicators, order books, and liquidation zones to make their decisions. I realized that many of these factors could be measured mathematically. With AI technology advancing rapidly, I decided to build a Bitcoin price prediction AI—essentially a neural network that could handle all this data and provide more accurate predictions.

I began this project by signing up for ChatGPT, where I asked if it could help me build such a model. While it was encouraging, I quickly discovered that it often agrees optimistically and forgets things a lot. Fortunately, with almost 20 years of coding experience, we were able to go back and forth to get it done. Although it took much longer than expected, the satisfaction of seeing it work has been worth the countless hours I’ve put in.

Building the Neural Network

Now, I have a working Bitcoin prediction neural network. It takes multiple inputs from various data sources and processes them using machine learning techniques. The next step is to connect this model to an interface and package it into an executable file. This way, I can start distributing the software and gathering feedback from users.

Technology and Indicators

The neural network relies on several key data sources and indicators:

Order Book Data: Provides insights into market support and resistance levels by analysing top bid and ask prices, the bid-ask spread, volume at these levels, and cumulative bid and ask volumes. These factors help the model gauge where liquidity is positioned and how it influences price movements.
Technical Indicators: The model incorporates a variety of widely used metrics, including:
- Moving Averages (MA7 and MA14): Short-term and medium-term moving averages that help identify trends and momentum.
- RSI (Relative Strength Index): A momentum oscillator that measures the speed and change of price movements, indicating potential overbought or oversold conditions.
- Bollinger Bands (UpperBand and LowerBand): These bands, based on price volatility, help identify price levels where the market may reverse or continue its trend.
- Stochastic Oscillator (%K and %D): A momentum indicator that compares a particular closing price to a range of prices over a certain period, providing signals of market trends and potential reversals.
- Volume: Measures the amount of trading activity to detect increases or decreases in market participation, often preceding price movements.
- MACD (Moving Average Convergence Divergence): A trend-following indicator that highlights the relationship between two moving averages of a security’s price.

Volatility: The model monitors market volatility, which is a critical input for assessing the degree of price movement. By analyzing volatility, the model can anticipate potential sharp price swings or periods of relative calm.
Fear & Greed Index: This index measures market sentiment by evaluating various factors such as volatility, trading volume, and social media activity. The Fear & Greed Index ranges from extreme fear to extreme greed, offering insights into whether market participants are overly optimistic or fearful. Integrating this index helps the model account for sentiment-driven movements, which can influence price, particularly during periods of intense market emotion (e.g., panic selling or euphoric buying). This is crucial for predicting price action during volatile or emotionally driven market phases.
Liquidation Zones: Analyzing liquidation zones is crucial for understanding how exchanges and market makers manipulate the market. By observing these zones, the model can anticipate where large clusters of leveraged positions are likely to get liquidated, influencing price action. Tools like Coinglass help visualize these areas, providing the neural network with data on how the price might be driven to target these zones.

The system is infinitely configurable, allowing users to adjust parameters such as the data decay rate, the number of data points used per prediction, and which indicators are active. This flexibility is essential because different setups may be more effective under various market conditions, and users can experiment with different combinations to optimize their trading strategy.

The Future: Automating State Switching

The first AI model I’m developing has three configurable states: bullish, bearish, and neutral. Users can manually switch between these states depending on market behavior. My plan for the future is to build a second AI that will automate this process. This AI will use a high-frequency data feed (updating every second) and liquidation zone data to decide which state to maintain.

Observing Manipulation

From my research, I’ve noticed that market makers and exchanges, acting like casinos, drive the price to hit these liquidation zones. Leveraged traders, who might be using high ratios like 50 to 1 or even 100 to 1, often cluster together, creating large liquidity zones. These zones can reach billions of dollars, and when they become too high, the market makers often drive the price through these levels to trigger liquidations.

The Approach

I want my AI to recognize these patterns and help avoid getting caught in the traps set by market manipulation. By letting the AI do the heavy lifting—analyzing the data, order books, and liquidation zones—I can make more informed decisions. It’s like having a sophisticated indicator that provides predictions based on mathematical models rather than emotions or market psychology. I don’t plan on relying solely on this model for trading; instead, I’ll use it alongside my intuition and only make trades when the model’s predictions align with my analysis.

Creating a Community

The first step is to finish the base layer, which includes building the interface and packaging the model. Once that’s polished, I plan to create a community where users can share configurations, give feedback, and report bugs or improvement suggestions. This feedback loop will be essential for refining the tool.

The Second AI: Predicting the Future State

Once I gather enough data and feedback from the first model, I’ll work on the second AI, which will automate state switching. This model will combine one-second data feeds and liquidation zone information to maintain the appropriate state, even during fake-outs or sudden price changes.

By building this sophisticated system, I aim to stay ahead of market manipulation. The second AI will detect patterns such as fakeouts and use liquidation zone data to avoid getting caught in traps set by market makers. My ultimate goal is to create a tool that not only provides accurate predictions but also helps traders navigate the complex and manipulated world of Bitcoin trading.

Conclusion

The journey so far has been challenging but rewarding. From my initial experiments with leverage trading to developing a neural network, I’ve learned a lot about the Bitcoin market. The first AI is just the beginning. With the future development of the second AI, I hope to create an advanced system capable of adapting to market conditions in real time. I look forward to building a community around this tool, refining it with user input, and taking it to the next level.